Who was Roman Mazurenko?
Roman Mazurenko was born in Belarus, the only child of an engineer and a landscape architect. As a teenager he spent a year on an exchange programme in New Mexico, then studied computer science in Dublin before moving to Moscow in 2007.
In Moscow he became a central figure in the city’s creative scene. With two friends, Dimitri Ustinov and Sergey Poydo, he founded a collective called Idle Conversation that ran magazines, festivals and club nights. Later he launched Stampsy, a tool for building digital magazines.
Eugenia Kuyda met him in 2008, when she was a young magazine editor writing about his collective. They became close friends and, as both turned to start-ups, each other’s main adviser. Kuyda co-founded Luka, an artificial intelligence company, and moved it to San Francisco in 2015. Mazurenko followed after a period in New York.
On 28 November 2015, while back in Moscow to finish visa paperwork, Mazurenko was hit by a speeding car at a pedestrian crossing. He died in hospital. Kuyda happened to be in Moscow that day.
Why a chatbot, of all things?
In the weeks after his death, friends suggested a photo book or a memorial website. None of it felt right to Kuyda. Mazurenko had used social media very little, and his body had been cremated, so there was no grave to visit. What remained were thousands of text messages.
Kuyda’s company already had the tools. Luka, backed by the start-up accelerator Y Combinator, made a messaging app for talking to bots, starting with restaurant recommendations. Google had released its TensorFlow machine-learning library as open source two weeks before Mazurenko died, and Luka was using it.
There was also a personal echo. In 2015 Mazurenko had applied to Y Combinator with an idea for “memorial forests”, a new kind of cemetery, and he wrote about people wanting to manage their digital estate. He was interested in the future of death long before anyone built a bot of him.
Kuyda watched the Black Mirror episode “Be Right Back” after his death and, according to The Verge, felt mixed about it: memorial bots seemed both inevitable and dangerous. Our page on what in “Be Right Back” is real now looks at that episode in detail.
How the Roman bot was made
Casey Newton’s long report for The Verge, published in October 2016, describes the process step by step:
| Step | What happened |
|---|---|
| Gathering | Kuyda asked friends and family, gently, for their messages. Ten people, including his parents, agreed. Together they shared more than 8,000 lines of text. |
| Curating | Kuyda copied hundreds of her own exchanges with him from the Telegram app and removed a handful she felt were too personal to share. |
| Base model | Luka trained a Russian-language neural network on more than 30 million lines of general Russian text. |
| Personal tuning | The network was tuned to use Mazurenko’s own words whenever it could, and to fall back on generic Russian when it could not. Only a small share of answers were his real phrases. |
| Launch | On 24 May 2016 Kuyda announced the bot on Facebook. Anyone with the Luka app could add @Roman and chat in Russian or English, or tap menu buttons to learn about his life. |
| Upgrade | Later in 2016 the bot moved from a “selective” model, which matched his real messages to questions, to a “generative” one that could recombine fragments into new sentences. |
That last change is the same trade-off every griefbot still faces. Picking real messages keeps the words authentic but limits what the bot can say. Generating new sentences makes it more fluent, but it can also say things the person never said.
How friends and family reacted
Reactions were divided, and the article reported them honestly. Mazurenko’s mother defended the project publicly and said that reading his thoughts helped her get to know her son better. His father, who has a technical background, was more reserved: he knew it was only a program, and it sometimes answered wrongly.
Four friends told Kuyda they found the bot disturbing and refused to use it. A former colleague criticised it in a public comment as rushed and half-finished. Other friends were struck by how much the short replies sounded like Roman’s texting style.
Hundreds of people tried the bot. With permission, Kuyda read anonymised logs and found that most people were not testing it. They were telling him they missed him, or saying things they never had time to say. In her own launch post she was modest about what it was, calling it “a shadow of a person”.
From memorial bot to Replika
The Verge reported in 2016 that Kuyda was steering Luka towards a new bot called Replika: part diary, part assistant, which asks about you and gradually learns your texting style. Replika was released in 2017 and became a companion app. Unlike the Roman bot, it is not built from a real person’s messages; each user shapes their own AI friend.
Replika has since faced scrutiny. In 2023 Italy’s data protection authority ordered it to stop processing Italian users’ data, citing risks for minors and vulnerable people, and in 2025 the authority announced a fine against its developer, Luka Inc. In early 2025 Kuyda stepped down as chief executive to start a new company, Wabi, and remained an adviser. Our comparison of Replika and Character.AI covers the app as it is today.
What the Roman bot still teaches
- Curation is part of consent. Kuyda asked before collecting messages and removed some she felt were too private. Every later project faces the same choice.
- People show different sides to different people. Newton noted that a bot built from everyone’s chats may reveal sides of a person that some friends were never meant to see.
- Grief responses vary. The same bot comforted his mother and upset several friends. Neither reaction is wrong.
- A bot is mostly a place to talk. Many users needed somewhere to say goodbye more than a convincing reply.
The story is one chapter of a longer history. See our timeline of griefbots and digital clones, and compare it with James Vlahos’s Dadbot, built with his father’s agreement, and the Project December story, which showed what large language models changed.
If you are thinking about something similar
Memory Clone is a Windows app, currently in alpha, that builds a private AI simulation from chats you export yourself, such as WhatsApp or Messenger. Like Kuyda’s project, it starts from real messages. Unlike it, the originals and the memory index stay on your PC, it learns style separately for each relationship, and every clone is presented as an AI simulation, not the person. It is text only and does not use voice notes, photos or video.
There is no right time, and choosing not to build anything is a good choice too. If you do, our guide to talking to a loved one who died with AI may help you decide, and you can read more on the download page.
If you are in crisis or thinking about harming yourself, contact your local emergency number now. You can also reach: UK & Ireland Samaritans, 116 123; United States 988 Suicide & Crisis Lifeline, call or text 988; elsewhere find a free, confidential helpline at findahelpline.com.
자주 묻는 질문
When did Roman Mazurenko die?
He died on 28 November 2015 in Moscow, after being hit by a car at a pedestrian crossing.
How many messages were used to build the Roman bot?
According to The Verge, ten friends and family members shared more than 8,000 lines of text. A general Russian language model trained on far more text filled the gaps.
Is Replika a copy of Roman Mazurenko?
No. The Roman bot inspired the company’s direction, but Replika is a companion app that each user shapes through their own conversations. It is not built from Roman’s messages.
Did Black Mirror inspire the Roman bot?
The Verge reported that Kuyda watched “Be Right Back” after Mazurenko’s death and had mixed feelings about it. Later retellings often describe the episode as an inspiration.
Can people still talk to the Roman bot?
We could not confirm that it is publicly available today. In 2022 Kuyda told MIT Technology Review that she still talks to it, especially around his birthday and the anniversary of his death.
출처
에 확인했어요. 타사 제품과 도움말 페이지는 자주 바뀌니 최신 정보는 링크를 확인하세요.
- Speak, Memory: When her best friend died, she rebuilt him using artificial intelligence (Casey Newton, 2016) (The Verge)
- Technology that lets us “speak” to our dead relatives has arrived. Are we ready? (Charlotte Jee, 2022) (MIT Technology Review)
- AI: the Italian Supervisory Authority fines company behind chatbot “Replika” (2025) (European Data Protection Board)
- Replika’s founder answers the critics (Tasnim Rödder, 2026) (MindSite News)
- Hollanek & Nowaczyk-Basińska, “Griefbots, Deadbots, Postmortem Avatars” (Philosophy & Technology, 2024) (Springer)
This article is general information, not medical, legal or financial advice. Memory Clone creates an AI simulation from the material you provide; it is not the person.